Anonymous advice: If you want to reduce AI risk, should you take roles that advance AI capabilities?

Abstract

This article discusses the complex question of whether individuals aiming to reduce risks from advanced artificial intelligence (AI) should consider roles that also advance AI capabilities. Capabilities research, focused on improving AI performance, and safety research, aimed at mitigating AI risks, are often intertwined. While some argue that advancing capabilities accelerates the timeline to potentially dangerous AI and thus detracts from safety efforts, others contend that capabilities work can be beneficial for safety in several ways. These include allowing more people concerned about AI risk to gain relevant skills, fostering collaboration and norm-setting within AI labs, enabling new safety research directions (e.g., interpretability, robustness), and offering valuable insights into the nature of AI systems. The impact of capabilities work depends on various factors, including the specific type of work, how it is used and disclosed, and the actor’s influence. Publicly disclosed research directly contributing to key advancements towards artificial general intelligence (AGI) is generally viewed as potentially harmful, whereas work on less central capabilities or alignment of existing models could be beneficial. The article emphasizes the importance of considering the potential for value drift when working in capabilities-enhancing roles and encourages individuals to engage with the AI safety community and literature to mitigate this risk. – AI-generated abstract.